Earthquakes are highly destructive phenomena that can result in fatalities and destruction of property and indirectly affect a country’s economy. The main objective is to improve the accuracy of magnitude estimation from crucial first 3 s of the earthquake record in early warning systems (EWS). Traditional empirical relations have limitations in providing precise earthquake magnitude predictions. MLMagPred addresses this by utilizing the sanitized grey wolf optimization (SGWO) algorithm for hyperparameter tuning of the XGBoost algorithm. The model inputs parameters from the early three seconds of preprocessed recordings obtained from a single seismic station. Typically, the occurrence of large earthquakes above MJMA ≥ 6 is low but causes more damage. Therefore, to counteract the drawback of the limited dataset, six engineered features include \({\tau }_{c}\) , Pd, Pv, \({\tau }_{p}\) , PIv, and Tva. The study focuses on 83 moderate to major earthquakes in Japan, totaling 18,330 records obtained from the Kyoshin network (K-NET) installed in Japan. These total records are divided into independent train, test, and validation sets comprising 12,989, 3740, and 1601 records, respectively. Mean absolute error (MAE) metrics for MLMagPred, the \({\tau }_{c}\) method, and the Pd method are observed to be 0.15, 2.23, and 1.76 MJMA, respectively. MLMagPred demonstrates superior performance compared to the widely used empirical relationship based on the \({\tau }_{c}\) and Pd methods proposed by Jin et al. [10]. Furthermore, a case study involving earthquakes with magnitudes of 7.3 (Mw) and 7.1 (Mw) illustrates the improved performance of MLMagPred in multi-station magnitude prediction. For instance, in the case of a 7.3 (Mw) earthquake that occurred on 11-03-2011 (Origin time 15:26), MLMagPred yields multi-station magnitudes of 7.2 while the \({\tau }_{c}\) and Pd methods predict the magnitude of same earthquake as 4.7 and 5.0 (Mw), respectively. Similar improvements are observed for a 7.1 (Mw) earthquake on 13-02-2021 (Origin time 23:08), where MLMagPred predicts multi-station magnitude of earthquake as 7.2 while the \({\tau }_{c}\) and Pd methods predict the magnitude of same earthquake as 4.5 and 4.8, respectively. The findings suggest that the utilization of metaheuristic algorithms can enhance the prediction accuracy of hyperparameters in XGBoost compared to grid search for earthquake magnitude prediction. The results of both single and multi-station studies underscore the potential of MLMagPred to improve large earthquake magnitude prediction accuracy in EWS.